Using process-oriented model output to enhance machine learning-based soil organic carbon prediction in space and...

Area of Science:

• Soil Science
• Environmental Science
• Data Science

Background:

• Soil organic carbon (SOC) dynamics are vital for climate change research and policy.
• Machine learning (ML) excels at spatial soil mapping but struggles with temporal dynamics.
• Process-oriented (PO) models capture temporal SOC changes mechanistically.

Purpose of the Study:

• To develop and test a hybrid model integrating PO and ML for space-time SOC stock prediction.
• To improve the accuracy and physical plausibility of SOC mapping.
• To support soil management and policy decisions under climate change.

Main Methods:

• Developed a hybrid model combining PO and ML for topsoil SOC stock prediction.
• Used PO model predictions as training data for the ML model in unsampled years.
• Employed a weighting parameter to balance PO model outputs and real measurements.

Main Results:

• Temporal trends from PO and hybrid models were similar, unlike the ML model alone.
• The hybrid model achieved the best performance with an RMSE of 0.29 kg m⁻².
• Demonstrated a 19% improvement in prediction accuracy compared to the ML model.

Conclusions:

• The hybrid framework enhances space-time soil carbon mapping accuracy and physical plausibility.
• This integrated approach offers valuable insights for soil management strategies.
• The model provides a robust tool for addressing climate change and human impacts on soils.